Professional Documents
Culture Documents
Williams Experimental Designs Balanced For The Estimation Of...
Williams Experimental Designs Balanced For The Estimation Of...
Summary
Where an experiment can be carried out by applying different treatments in
succession to the same unit of experimental material, accurate comparisons can be made
between the effects of different treatments. To allow for the residual effect of previous
treatments on the result obtained for any given treatment, it is desirable to adjust the
results for such effects. Methods of constructing balanced designs for the estimation
of these residual effects are described in this paper, and are summarized as follows.
Designs balanced for effect of single preceding treatment : Wnen n , the number of
treatments, is even, a balanced design is possible with n replications ; when n is odd,
2% replications are required.
Designs balanced for the effects of any number of preceding treatments, ignoring the
znteraction of residual effects : When n is a prime or a power of a prime, a balanced
design is possible in n(n-1) replications, which may be set out as a set of n-1 mutually
orthogonal Latin squares, with the same first columns. Designs which are not expressible
a s mutually orthogonal Latin squares are also possible.
Desigm balanced for the eifeat of the two preceding treatments and their interactions :
A design can be developed from a set of n-l mutually orthogonal Latin squares obeying
certain restrictions.
The method of analysis of designs of this type is set out in detail, together with a
numerical example. Direct effects of treatments are shown to be only slightly con-
founded, the maximum confounding being 4 per cent., when there are three treatments.
These designs have u-ide applicability wherever successive treatments can be
applied to the same unit of experimental material.
I. INTRODUCTIOX
Experiments carried out to compare the effects of different treatments on
a given material are discussed in this paper. It deals particularly with those
cases in which the different treatments can be applied in succession to the same
unit of the material. Where this can be done, the treatment effects will be
compared as accurately as possible, since the variation between units will not
affect the comparisons between the effects of treatments. For this accuracy to
be fully realized, however, it is important to make allowance for the effect of
previously applied treatments on the response of the unit of material to a given
treatment.
The estimation of the effects of previous treatments is of direct interest in
experiments, such as crop rotation trials in agriculture, where the aim is to
determine the succession of treatments (i.e. crops, fallow, etc.) over a number
of years, which will give the best results for the series taken as a whole. This
applies, whether the criterion is yield, market value, or the continued fertility
of the soil. On the other hand, such estimation is of indirect interest only, in
experiments such as those on the feeding of animals. In such cases, while a
group of aninlalx might be tested with a succession of treatments (rations),
the primary interest centres In determining the particular ration which gives
best results according to some criterion. The effect of previous treatments is of
indirect interest, and is estimated only as a correction in determining the effect
of the treatment which follows.
The method of carrying out an experiment by applying successive treatments
to the same unit of material has wide applicability. The main limiting factors
are (i) the length of time taken, compared with simultaneous application of the
different treatments to different units of material, and (ii) the suitability of the
material to repeated treatments. Por example, if the treatment were destructive,
the method would not be applicable.
The unit of experimental material to which successive treatments are
applied will be called a site. The effect of the treatment being applied will be
called the direct effect, and the effect of previous treatment the residual effect,
following Cochran, Autrey, and Cannon(1). We shall discuss the design of
experiments for the estimation of the residual effect, firstly, of the immediately
preceding treatment, and secondly, of the two immediately preceding treatments.
The analysis of such experiments will be discussed and exemplified only for the
former case.
D
Other sets of suitable initial rows have been determined by trial, and are
given for n=4, 6, 8, in Table 1. When one initial row has been found, others
can be derived from it by multiplying each element by a residue which is prime
to rz, and adding an arbitrary integer. Such rows are equivalent, since the
effect of multiplication is the same as permutation of the numbers representing
the treatments, and a rearrangement of the rows, and the addition simply
permutes the rows. Only a standard representative of each such set of rows is
therefore tabulated.
Case I1 : n Odd
When n is odd, the sum of the residues is a multiple of n. Hence, if every
difference were to occur once in a row, the last element would be the same as
the first. It can be seen, therefore, that balanced designs based on a single
cyclic Latin square are impossible. I t is, however, possible to achieve balance
with a pair of such squares, the differences not represented in one square being
represented twice in the other.
For example, when n=5, a possible pair of sets of differences, generating
s pair of Latin squapes which together provide balance, is
1 2 1 3 ' 2 4 3 4
154 E. J . WILLIAMS
Paire of initial rows are given, for m=3, 5 , 7 in Table 2. As for even values
of lz, other pairs may be derived from these by multiplying each element by a
residue prime to n, and adding an arbitrary integer.
ESTLMATION O F RESIDUAL EFFECTS 155
* Since this paper went to press, designs based on sets of oyclic Latin squares which are not
orthogonal have been found, including designs for n=6 and n= 10. I t is intended t o describe
these in a subsequent note.
ESTIMATIOX OF RESIDUAL EFFECTS 157
Classification
Row Symbols Square A Square B Square C
Column .. I I1 111 I V I I1 111 I V I I1 I11 I V I I1 I11 I V
1 1 1 1 1 4 2 3 1 3 4 2 1 2 3 4
2 2 2 2 2 3 1 4 2 4 3 1 2 1 4 3
3 3 3 3 3 2 4 1 3 1 2 4 3 4 1 2
4 4 4 4 4 1 3 2 4 2 1 3 4 3 2 1
and the corresponding initial row of the first Latin square of the set is
The m(n-1) rows obtained by multiplying the elements of the initial row by
1, 2, 3, . . . . . . n-1, and in each case adding to the resulting elements
0, 1,2, . . . . . . n-1 give the required design balanced for pairs of residual effects.
This proves the existence of such a design when n is any odd prime.
This general solution is the same as that found above for a single Latin
square giving balance for the effect of one preceding treatment only, when
n is even, or for the pair of Latin squares required when n is odd.
Other initial rows also give balanced designs in the same way ; these are
tabulated in Table 3, for 12=3, 5, 7.
Case I1 : n a Power of a Prime
When n=pk, where p is prime and 3c an integer, each treatment may be
represented by one of the elements of a Galois field. Galois fields exist only for rz
a prime or a power of a prime. A full treatment of the properties of these fields
will be found in Carmichael (4, pp. 242-88).
Every field has a primitive element a, with the property that the equation
m m = l is satisfied only by m a multiple of n-1. Consequently, the elements
x, a2, x3, . . . . . . an-l (=I)are all distinct, and therefore include every non-zero
element of the field.
It can be shown that such s primitive element satisfies an equation
P(m)=0, where P(x) is an irreducible polynomial of degree 5, whose coefficients
are members of GP(p). P ( m ) is called a minimum function. It follows that
every element of the field may be expressed as a polynomial in x of degree k-1,
namely
+
c,, +c1x . . . . . . +C k-lmkl,
160 E. J. WILLIAMS
. ...
where the coefficients c,, c,, . . c,-, are members of Gl?(p). Addition and
multiplication are defined as ordinary addition and multiplication, followed by
simultaneous reduction modulo p and modulo l?(x).
We note that, when n is the power of an odd prime,
xn-1-1 =O
but
therefore -
When n is a power of 2,
1+1=o.
We are now in a position to adapt the procedure used when n is prime, to
the derivation of a balanced design when .n is a power of a prime. We require
to find a set of 12-1 non-zero differences among the elements of the field, such
that (i) no cumulative sum of successive differences is 0, and (ii) the ratios of
12-1
successive differences take all values except 0, x 7 (when n is odd) and all values
except 0, 1 (when is a power of 2).
As for the case of n a prime, the first and lest differences must be equal.
..
If the set of differences be multiplied by 1, x, x2, . . . . in turn, n -1
sets will be generated. Corresponding to each such set of differences, a Latin
square can be formed. The argument given above for the case when n is prime
can be extended to show that the Latin squares so constructed are mutually
orthogonal, and that every ordered trio of elements is represented once. This
set of squares therefore gives the required balanced design.
Although it is believed that a Bolution exists for every n which is a power of
a prime, no general solution has been found. Solutions for the particular cases
likely to be required in practice, namely, n=4, 8, 9, are, however, presented in
Table 4. The initial row is given, also the minimum function and the power
cycle for the primitive root x. The initial row is standardized by choosing 0 and 1 -
as the first two elements.
From any given solution, other solutions may be derived by raising each
element to a power whose index is a power of p. The addition and multiplication
relationships among the new elements are exactly the same as those among the
old. For the pth power of a sum is congruent (modulo p ) to the sum of the
pth powers ; while the pth power of a product is obviously identical with the
product of pth powers. There are just k different solutions so derivable ; for
if x is any element,
$9 =x" =x.
This fact enables initial rows to be further standardized. For example, when
n=9, the following two initial rows are equivalent :
0 1 x $5 x2 $3 $4 $7 x6
0 1 x3 m7 x6 x rC4 x5 x2
In Table 4, the solution is given for which the third element is the lowest possible
power of x.
ESTIMATION O F RESIDUAL EFFECTS 161
As an example of the generation of a balanced design from the initial row
.
given in Table 4, the complete design. is given for n =4. I n setting out the design
in this way, it-is convenient to represent the zero element by 0, the unit element
..
by 1, and the successive powers of m by 2, 3, . . . . n-1.
0 1 2 3 0 2 3 1 0 3 1 2
1 0 3 2 1 3 2 0 1 2 0 3
2 3 0 1 2 0 1 3 2 1 3 0
3 2 1 0 3 1 0 2 3 0 2 1
DESIGNS BALANCED FOR THE EFFECTS OF PALRS OF RESIDUAL EFBECTS. INITIAL ROWS FOR
. GENERATING SETS OF ORTHOGONAL SQUARES WHEN ?% IS A POWER OF A PRIME
12
-.-
Initial ROW 1 fE2E
I
I
I
Power Cycles
p =A.
nP.-G
mn2
The estimates given here take into account only the information given by
comparisons within sites (actually, within groups of sites with the same final
treatment). If the information provided by comparisons between groups of
sites be taken into account, slightly more accurate estimates result.
It can be shown that residual effects are confounded 1/(n2--n)with groups
of sites and with treatments. Since only that part confounded with sites is
potentially recoverable, the fraction of recoverable information confounded with
sites is l / ( m 2 -n -1).
I n the same way it is found that direct effects of treatments are confounded
l / ( n 2-n) with residual effects when sites are not taken into account. The fraction
thus confounded is of course not recoverable, so that the fraction potentially
recoverable is (n2-.n -1)/(n2 -n). When sites are taken into account, an
additional fraction l / ( ( n 2-n)(n2 -n -1)) is confounded. Hence, the fraction of
potentially recoverable information on direct effects, confounded with sites, is
l / ( n 2-n The efficiency factors for estimates from comparisons within
sites are therefore as follows :
Residual effects
1 Residual
These factors are presented, for small values of ?z, in Table 5. Unless n is small,
the gain due to recovery of information from comparisons among sites will be
slight, and will moreover be offset to some extent by the loss of information due
to inaccuracies in weight,ing. However, recovery of information on residual
effects may sometimes be worth while, when n is 3 or 4.
I n many cases, where the estimation of the residual effect is not of direct
interest, and it is required only as an adjustment to the average effect of a
164 E. J. WILLIAMS
treatment, estimates of the residual effect with improved accuracy are irrelevant.
The recovery of information Is, however, discussed below, to meet the oases in
which the residual effects are of direct interest.
Ordinarily, two analyses of variance will be required for testing the
significance of treatments and of the djustment for residual effects. The sum
of squares for direct effects, corrected for residual effects, and the sum of squares
for residual effects, corrected for direct effects, will both need to be calculated.
The two analyses are set out in Table 6.
Sum of Squares
Degrees of
Freedom Direct Effects Residual Effects
Adjusted / Adjusted
Sites-Between groups
Within groups ..
Positions .. ..
Direct effects .. ..
Residual effects ..
Error .. .. .. by subtraction
Total
where E is the error mean square from the analysis of variance. When the
adjustment for residual effects is made, the variance of a difference is increased to
ESTIMATION OF RESIDUAL EFFEOTS 165
The variance of the difference between the estimates for two residual
effects is
2nE
m(n2-n -2) '
( d ) Recovery of Information between Groups of Bites
Let the error variance be 1/W, and the variance between sites of the same
group, on a single-result basis, be l/W'.
I f E is the error mean square and S the mean square between sites within
groups, their expected values are respectively 1 / W and l/Wr. However, if E
exceeds B, we assume W and W' to be equal, and take 1/E as the eatimate of
each.
The estimate of the residual effect of treatment, after including information
from groups of sites, is
The expression for the variance of differences may be compared with that obtained
above, without recovery of information. The expression in brackets represents
the reciprocal of the relative efficiency with recovery of information.
Replication
Numerical Example
The example given below is taken from experimental pulp evaluation
technique. Samples of pulp suspension at varying concentrations are beaten
in a Lamp& mill, to determine the effect of concentration on the properties of
the resulting sheets. Observations of the condition of the mill after each beating
indicate that certain concentrations of pulp have m effect on the mill which may
166 E. J. WILLIAMS
affect the next besting. I n order to allow for this effect if it exists, a balanced
design is adopted. There are six concentrations used, in all, and six replications
are carried out. The layout of the experiment is shown in Table 7 and the results
obtained for the burst factor of the resulting sheets in Table 8.
TABLE8
BURST FACTOR RESULTS F O R THE EXPERIXENT O F TABLE 7
Replication Total
Total
TABLE9
TOTALS AND SUMS O F S Q U m E S REQLTRED FOR ANALYSIS
In terms of
estimates
In terms of
totals
i
1
2
3
4
5
6
Add P ,
I
I
Divisor for sum of squares .. , 216
Sum of squares . . .. .. 69.298 24.076 1 23.436
I
ESTIi?IATION O F RESIDUAL EFFECTS 167
For numerical work it is convenient to use the quantities given below in cal-
culating the estimates and the sums of squares in the analysis of variance.
Corresponding to any treatment, all but the first of these quantities may be
TABLE10
ANALYSIS O F VARIANCE
Sum of Squares
Degrees
of Mean Square
Freedom Direct Effects Residual Effects (Adjusted)
Adjusted Adjusted
Replications ..
Positions in order
of beating ..
Direct effects . .
Residual effects
Error .. ..
Total .. 165.062 I
1
I
TABLE11
ADJUSTED TREATMENT MEANS AND ESTIMATES O F RESIDUAL EFFECTS
I .
Adjusted Residual
Treatment 1 Mean Effect
1
2
3
4
5
6
Standard error
of difference 0.46
168 E. J. WILLIAMS
Divisor
A =mn2t', =nT -G ma3
B i =nan(n2-n-l)r', =n2Ri +%Pi +nPl -(n +1)G rnn3(n2-n -1)
C i =mn(n2 -n -2)ri =n2Ri +nPi+%Ti +nPl - ( n f 2 ) G mn3(n2-n - 2 )
=B, + A i
D , =mw(n2-n -2)ti =(n2 -n -1)l', +nR, + P i +PI -nG mn(n2-n -1)(n2 -n -2)
The detailed analysis of the results of the experiment is set out in Tables
9, 10, and 11.
V. ACKNOWLEDGMENT
The work described in this paper was carried out as part of the research
programme of the Section of Mathematical Statistics, 0.8.I.R.O.
VI. REFERENCES
(1) COCHRAN, W. G., AUTREY, K. M., and CANNON, C. Y.-A double change-over design for dairy
cattle feeding experiments. J. Dairy Sci. 24 : 937 (1941).
(2) B o s ~ R.
, C.-On the application of the properties of Galois fields to the problem of con-
struction of Hyper-Graeco-Latin squares. Sankhyd 3 : 323 (1938).
(3) STEVENS, W. L.-The completely orthogonalized Latin square. Ann. Eugen. 9 : 82 (1939).
(4) CAR~CHAEL, R. D.-" Introduction to the Theory of Groups of Finite Order." (Ginn and
Co. : Boston and London, 1937.)